[Core] gpu memory scheduling prototype

[jonathan-anyscale]

A prototype to allow user to specify _gpu_memory as alternative to specify fractional gpu to the remote function. The field _gpu_memory defined as "The gpu memory request in megabytes for this task/actor from a single gpu, rounded down to the nearest integer.".

# Assume we have a single GPU with 1000mb total memory

@ray.remote(_gpu_memory=100 * 1024 * 1024) # allocate 100mb from a single gpu to this task
def task():
    # contains { 'GPU': [assigned_gpu_id]}
    resources = ray.get_runtime_context().get_resource_ids()

@ray.remote(_gpu_memory=1001  * 1024 * 1024) # task can't be scheduled due to requested memory > total gpu memory
def expensive_task():
    resources = ray.get_runtime_context().get_resource_ids()

@ray.remote(num_gpus=1, _gpu_memory=100  * 1024 * 1024) # task can't request both num_gpus and gpu_memory
def not_allowed_task():
    resources = ray.get_runtime_context().get_resource_ids()

Implementation detail: _gpu_memory will be converted to num_gpus before being scheduled where num_gpus = _gpu_memory / gpu_total_memory and we check if the 0 <= num_gpus <= 1 representing fractional GPU request.

Implementation Detail

_gpu_memory is an alternative representation of num_gpus where num_ gpus = _gpu_memory / node_total_ gpu_ memory where node_total_gpu_memory is total gpu memory of the GPU type in the node.

Thus, we convert gpu_memory to GPU resource when scheduling depending on what GPU type the scheduled node has and update GPU resource value stored in NodeResources. Additionally, since GPU has precision of $10^{-4}$, we rounded up the converted gpu_memory resource with the GPU precision.

Status

Currently working on cluster with multi-nodes setup, but not working with autoscaler yet.

Why are these changes needed?

Related issue number

Closes #37574

Checks

• I've signed off every commit(by using the -s flag, i.e., git commit -s) in this PR.
• I've run scripts/format.sh to lint the changes in this PR.
• I've included any doc changes needed for https://docs.ray.io/en/master/.
  • I've added any new APIs to the API Reference. For example, if I added a
    method in Tune, I've added it in doc/source/tune/api/ under the
    corresponding .rst file.
• I've made sure the tests are passing. Note that there might be a few flaky tests, see the recent failures at https://flakey-tests.ray.io/
• Testing Strategy
  • Unit tests
  • Release tests
  • This PR is not tested :(